深度学习驱动的分子表征与生成：药物发现的新途径

需积分: 11 141 浏览量更新于2024-07-02 收藏 3.84MB PDF 举报

本文主要探讨了分子图表示学习与生成在药物发现中的应用，以深度学习特别是图神经网络(GNN)为基础，推动了该领域的创新。作者 Benson Chen 深入研究了分子表征和分子生成这两个关键问题，这两个方面对于药物发现至关重要。在分子表征部分，论文介绍了一种新颖的Transformer式架构，它突破了传统的图神经网络范式，提出了分子原型网络。这种架构能够捕捉分子结构的复杂性，并通过学习分子的性质和反应预测，有效地理解分子间的微妙关系。作者探索了如何利用这种新型网络来生成有意义的分子表示，这在药物设计中有助于识别潜在的活性化合物。在分子生成方面，作者关注的是如何生成具有多种性质的药物候选分子。他们采用了一种创新的方法，首先通过学习分子片段的分布词汇，构建一个化学空间的语言模型。这种方法允许系统在生成新分子时保持化学的可行性和功能性。随后，论文深入研究了如何有效地探索化学空间，以便发现具有优良药效的潜在化合物。总结来说，这篇论文不仅提升了分子表征的深度学习技术，还引入了创新的分子生成策略，对药物发现过程中的分子表示学习和生成技术进行了重大贡献。这种结合了深度学习和图神经网络的研究，有望加速药物研发的速度和效率，为未来的药物发现带来实质性的进步。

4.2 Prediction accuracies when tested on our template split of the USPTO

dataset, for which any template-based model would get 0% accuracy on

the test set. We see that our template-free methods can still generalize

to this test set. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4.3 The left table shows the comparison of number of unique reactions

predicted by the base model vs. the mixture model (holding other

factors constant). The right table shows human evaluation metrics,

in which a human was asked to rate whether the outputs of the base

model or the mixture model was more diverse, or neither. . . . . . . . 63

4.4 Prediction accuracies using a masked pre-training objective; numbers

for other models taken from table 4.1 for easy comparison. The model

trained with this masked pre-training object performs worse than the

baseline model, suggesting the importance of having chemically-relevant

pre-training tasks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

5.1 FaST outperforms all baselines on both single-property and multi-

property optimization. Error bars indicates one standard deviation,

obtained from averaging 5 random seeds. . . . . . . . . . . . . . . . 80

5.2 Results on the constrained penalized logP task. FaST signiﬁcantly

outperform all baselines. . . . . . . . . . . . . . . . . . . . . . . . . . 86

5.3 Results on the GSK3𝛽+JNK3+QED+SA task using AP ﬁngerprints

instead of Morgan ﬁngerprints for novelty/diversity computation. . . 88

Some of these ﬁngerprints like the structure-based MACCS [30] ﬁngerprints, are highly

speciﬁc representations, consisting of an indicator function for a ﬁxed set of predeﬁned

structures. Other ﬁngerprints, topological and circular, which include Morgan ﬁnger-

prints are more ﬂexible. These ﬁngerprints capture local topology by enumerating

paths or circular neighborhoods. However, the problem still lies within the deter-

ministic nature of the generating method: if these predeﬁned rules do not capture

the right representation for the task, they will not work well. For instance, property

cliﬀs, a phenomenon in which similar molecules exhibit diﬀerent properties, remain

a challenging problem for many small molecule problems. This problem is especially

poignant for molecular ﬁngerprints, because the featurization is ﬁxed. However, using

deep models is not a bandage for this problem either, as deep models can easily overﬁt

to the data, and oﬀer poor generalization.

Therefore, it is crucial that our deep learning models incorporate the right kinds

of structural biases. Graph neural networks operate through an iterative aggrega-

tion scheme, wherein at each step, nodes aggregate information from its neighbors.

Successively, a node should incorporate more and more information about a larger

neighborhood. The node representations are eventually aggregated into a single vec-

tor representing the graph. While sometimes eﬀective, this simple paradigm may not

always incorporate the right kind of biases for molecular tasks. For instance, this

local neighborhood aggregation can fail to capture long-range dependencies that are

important when considering properties of molecules. Even more so, perhaps aggrega-

tion on the 2-D molecular graph is not suitable for the ideal molecular representation,

and we should look at 3-D structures instead. There are many considerations for the

development of deep models for molecules, but they require the correct structure to

be eﬀective. Fingerprint representations are simple, but they are inﬂexible and often

involve a lot of human-engineered rules. Deep models, on the other hand, can easily

overﬁt and fail to capture the correct structural representations.

1.1 Machine Learning Applications for Drug Discov-

ery

With that in mind, I will next explain how machine learning has been used in drug dis-

covery, and my eﬀorts to tackle these problems using more chemically-inspired models.

Drug discovery is a process consisting of several stages, and machine learning is par-

ticularly suitable for early-stage drug discovery, searching for new drug molecules.

During the discovery phase, high throughput screening (HTS) is conducted on large

libraries of molecules, which yields candidate molecules, known as hits. These hit

molecules then undergo more screening and optimization to generate a smaller set

of lead molecules. Usually, the HTS ﬁlters are quite relaxed, only testing for gen-

eral applicability, while generating lead compounds require much more scrutinized

evaluation and optimization. Once the lead compounds have been further optimized,

promising candidates are then tested on animals and later human trials.

The selection of hit and lead compounds is the ideal frontier for machine learning

methods to pave new improvements. In order to search for hit compounds, biological

assays are conducted to assess the properties for a large library of molecules. Using

machine learning, we can perform virtual screening (VS), which is to use computa-

tional tools to predict the properties instead of performing actual assays. This allows

us to speed up the screening process as well as being able to screen a much larger set of

molecules. Prior to machine learning, QSAR methods were broadly applied to virtual

screening. In its most basic form, QSAR methods use a variety of hand-engineered

descriptors, such as simple features including atom and bond counts, molecular weight

and ring information; more complex descriptors include higher-order topological fea-

tures and physicochemical properties. However, as mentioned earlier, these methods

that rely on a lot of hand-engineered features can underﬁt by not capturing the correct

representations, or overﬁt by introducing many noisy features.

Once hits are discovered, they usually need to be optimized. This is a challenging

problem because drugs need to satisfy multiple constraints in addition to their bio-

logical activity: they have to have exhibit certain properties like water-solubility or

剩余103页未读，继续阅读

努力+努力=幸运

粉丝: 2
资源: 136

深度学习驱动的分子表征与生成：药物发现的新途径

M2GRL_A Multi-task Multi-view Graph Representation Learning Framework for Web-sc

图表示学习(Graph representation learning)-AAAI-19-Tutorial

Graph Representation Learning

Innovations in Graph Representation Learning.pdf

图表示学习（Graph representation learning） （part0 and part1）.zip

图表示学习（Graph representation learning） (part2 and part3).zip

2019-[斯坦福 Leskovec]-Hierarchical Graph Representation Learning w

predicting patient outcomes with graph representation learning

2019-Likelihood-Free Inference and Generation of Molecular Graph

欠采样算法matlab实现代码-Inductive-Graph-Representation-Learning-for-Fraud-Detec

最新资源

图表示学习（Graph representation learning）（part0 and part1）.zip